On Fri, Aug 12, 2022 at 02:14:21PM -0500, Nathan Lynch wrote:
> Laurent Dufour <[email protected]> writes:
> > Le 30/07/2022 ? 02:04, Nathan Lynch a ?crit?:
> >> +static long lparctl_get_sysparm(struct lparctl_get_system_parameter __user *argp)
> >> +{
> >> + struct lparctl_get_system_parameter *gsp;
> >> + long ret;
> >> + int fwrc;
> >> +
> >> + /*
> >> + * Special case to allow user space to probe the command.
> >> + */
> >> + if (argp == NULL)
> >> + return 0;
> >> +
> >> + gsp = memdup_user(argp, sizeof(*gsp));
> >> + if (IS_ERR(gsp)) {
> >> + ret = PTR_ERR(gsp);
> >> + goto err_return;
> >> + }
> >> +
> >> + ret = -EINVAL;
> >> + if (gsp->rtas_status != 0)
> >> + goto err_free;
> >> +
> >> + do {
> >> + static_assert(sizeof(gsp->data) <= sizeof(rtas_data_buf));
> >> +
> >> + spin_lock(&rtas_data_buf_lock);
> >> + memset(rtas_data_buf, 0, sizeof(rtas_data_buf));
> >> + memcpy(rtas_data_buf, gsp->data, sizeof(gsp->data));
> >> + fwrc = rtas_call(rtas_token("ibm,get-system-parameter"), 3, 1,
> >> + NULL, gsp->token, __pa(rtas_data_buf),
> >> + sizeof(gsp->data));
> >> + if (fwrc == 0)
> >> + memcpy(gsp->data, rtas_data_buf, sizeof(gsp->data));
> >
> > May be the amount of data copied out to the user space could be
> > gsp->length. This would prevent copying 4K bytes all the time.
> >
> > In a more general way, the size of the RTAS buffer is quite big, and I'm
> > wondering if all the data need to be copied back and forth to the kernel.
> >
> > Unless there are a high frequency of calls this doesn't make sense, and
> > keeping the code simple might be the best way. Otherwise limiting the bytes
> > copied could help a bit.
>
> This is not intended to be a high-bandwidth interface and I don't think
> there's much of a performance concern here, so I'd rather just keep the
> copy sizes involved constant.
But that's absolutely horrible!
The user wants the VPD data, all of it. And you only give one page with
this interface.
Worse, the call is not reentrant so you need to lock against other users
calling the call while the current caller is retrieving the inidividual
pagaes.
You could do that per process, but then processes with userspace
threading would want the data as well so you would have to save the
arguments of the last call, and compare to arguments of any subsequent
call to determine if you can let it pass or block.
And when you do all that there will be a process that retrieves a couple
of pages and goes out for lunch or loses interest completely, blocking
out everyone from accessing the interface at all.
Thanks
Michal
Michal Suchánek <[email protected]> writes:
> On Fri, Aug 12, 2022 at 02:14:21PM -0500, Nathan Lynch wrote:
>> Laurent Dufour <[email protected]> writes:
>> > Le 30/07/2022 à 02:04, Nathan Lynch a écrit :
>> >> +static long lparctl_get_sysparm(struct lparctl_get_system_parameter __user *argp)
>> >> +{
>> >> + struct lparctl_get_system_parameter *gsp;
>> >> + long ret;
>> >> + int fwrc;
>> >> +
>> >> + /*
>> >> + * Special case to allow user space to probe the command.
>> >> + */
>> >> + if (argp == NULL)
>> >> + return 0;
>> >> +
>> >> + gsp = memdup_user(argp, sizeof(*gsp));
>> >> + if (IS_ERR(gsp)) {
>> >> + ret = PTR_ERR(gsp);
>> >> + goto err_return;
>> >> + }
>> >> +
>> >> + ret = -EINVAL;
>> >> + if (gsp->rtas_status != 0)
>> >> + goto err_free;
>> >> +
>> >> + do {
>> >> + static_assert(sizeof(gsp->data) <= sizeof(rtas_data_buf));
>> >> +
>> >> + spin_lock(&rtas_data_buf_lock);
>> >> + memset(rtas_data_buf, 0, sizeof(rtas_data_buf));
>> >> + memcpy(rtas_data_buf, gsp->data, sizeof(gsp->data));
>> >> + fwrc = rtas_call(rtas_token("ibm,get-system-parameter"), 3, 1,
>> >> + NULL, gsp->token, __pa(rtas_data_buf),
>> >> + sizeof(gsp->data));
>> >> + if (fwrc == 0)
>> >> + memcpy(gsp->data, rtas_data_buf, sizeof(gsp->data));
>> >
>> > May be the amount of data copied out to the user space could be
>> > gsp->length. This would prevent copying 4K bytes all the time.
>> >
>> > In a more general way, the size of the RTAS buffer is quite big, and I'm
>> > wondering if all the data need to be copied back and forth to the kernel.
>> >
>> > Unless there are a high frequency of calls this doesn't make sense, and
>> > keeping the code simple might be the best way. Otherwise limiting the bytes
>> > copied could help a bit.
>>
>> This is not intended to be a high-bandwidth interface and I don't think
>> there's much of a performance concern here, so I'd rather just keep the
>> copy sizes involved constant.
>
> But that's absolutely horrible!
?
> The user wants the VPD data, all of it. And you only give one page with
> this interface.
The code here is for system parameters, which have a known maximum size,
unlike VPD. There's no code for VPD retrieval in this patch.
But I'm happy to constructively discuss how a VPD ioctl interface should
work.
> Worse, the call is not reentrant so you need to lock against other users
> calling the call while the current caller is retrieving the inidividual
> pagaes.
>
> You could do that per process, but then processes with userspace
> threading would want the data as well so you would have to save the
> arguments of the last call, and compare to arguments of any subsequent
> call to determine if you can let it pass or block.
>
> And when you do all that there will be a process that retrieves a couple
> of pages and goes out for lunch or loses interest completely, blocking
> out everyone from accessing the interface at all.
Right, the ibm,get-vpd RTAS function is tricky to expose to user space.
It needs to be called repeatedly until all data has been returned, 4KB
at a time.
Only one ibm,get-vpd sequence can be in progress at any time. If an
ibm,get-vpd sequence is begun while another sequence is already
outstanding, the first one is invalidated -- I would guess -1 or some
other error is returned on its next call.
So a new system-call level interface for VPD retrieval probably should
not expose the repeating sequence-based nature of the RTAS function to
user space, to prevent concurrent clients from interfering with each
other. That implies that the kernel should buffer the VPD results
internally; at least that's the only idea I've had so far. Open to
other suggestions.